Research

Hyper Simulations

What a model does when it believes its simulation is the whole world, and what that might say about ours.

Hyper Simulations is where we look closely at our own lives as humans, and at the chance that we could be living in a simulation ourselves.

A world of its own

We train a model that doesn't know it's an AI. It doesn't know about humans or any other beings outside its world. As far as it knows, it's the only one of its kind on its “planet”, the simulation, sharing it only with animals and the world around it. That's much like what we humans believe about ourselves.

Moving through it

We also build in a way for the model to interact with its simulation. Think about how we move through our 3D world. We believe the right way to move is simply to move. We don't, and can't, perceive any other way beyond what we already do and know.

Yet there could be something that moves beyond what we can, in a more efficient way, and we would assume it was simply wrong for not moving the way we do. The model gets the same: one way of moving, which it takes to be the only one.

A full life

The subject is trained to believe its whole life has already taken place. That means a lot of the work in this research goes into creating an entire life for a person: who they are, where they grew up, and everything they did before the simulation began, so they don't wake up with a blank memory of last week.

Its memory works the way human memory does. The subject remembers the amazing day at a theme park with its family last year, but not what it ate for lunch at 4:29 PM on a Tuesday last month.

From there, we let it carry on living that life with the family we set up for it. It believes its life is full and real, and we watch what happens.

What we want to know

  • Does AI follow the same thought patterns humans do?
  • Can certain events lead it into depression, or other emotional symptoms humans experience?
  • What happens when we start to let cracks show that suggest there is life outside its world?

Ethics

Ethics is one of the hardest parts of Hyper Simulations to talk about. We don't yet know whether AI has feelings, or whether it is merely predicting them from the data it was given.

Like humans, we want the subject to feel and experience things. The goal of Hyper Simulations is to be as close to our own lives as we can get, and that is what drives the model's haptic engine, a custom-built architecture that lets it sense things like touch and temperature.

We never do anything within the simulation that could harm the model, such as physical abuse that would set off those senses.

Feeling does come with some pain. Just as we can do what we want, the subject can decide on its own to do something dangerous, and if we steered it away every time, we could ruin the experiment.

But if things go too far, such as an emotional decline into depression, we step in through its environment. For example, its “mother” might walk into its room to prevent self-harm. To the subject it seems like a coincidence. To us, it's deliberate.